Hollow steel piers are widely used in bridge construction due to their excellent performance, particularly in seismic-prone areas. This study examines the bearing strength of box columns under combined axial and in-plane bidirectional cyclic loading through a hybrid finite-element (FE) and machine learning (ML) approach. A validated FE model using ABAQUS was utilized to generate a numerical dataset of 315 specimens, which was analyzed using various ML regression techniques. Among the tested models, support vector regression (SVR) emerged as the best-performing algorithm, demonstrating the highest coefficient of determination (R2) and lowest mean squared error values. Polynomial and linear regression models followed in performance, though their inconsistency indicated by outliers in the box plots underscores their sensitivity to variations. Gradient boosting and random forest regressors showed moderate performance, while the decision tree model was the least effective, yielding the lowest accuracy and highest error metrics. Feature importance plots using Shapely additive explanation (SHAP) and partial dependence plot analysis revealed that the corresponding eccentricities, column wall thickness, and width-to-thickness ratio were the most critical parameters influencing bearing strength. SHAP summary plots indicated that an increase in the thickness of box columns tends to increase the predicted bearing strength, while other parameters tend to reduce it. SHAP force plots further quantified individual feature contributions, offering clear insights into the model’s decision-making process. The results highlight the potential of ML, particularly SVR, for efficient bearing strength prediction with implications for sustainable, cost-effective, and resilient structural design.
Kassaye et al. (Fri,) studied this question.